3 files changed, 95 insertions, 26 deletions

diff --git a/llvm/lib/Target/AMDGPU/AMDGPUAlwaysInlinePass.cpp b/llvm/lib/Target/AMDGPU/AMDGPUAlwaysInlinePass.cpp
index c27425443ab..d4bbb2c1eb8 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUAlwaysInlinePass.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUAlwaysInlinePass.cpp
@@ -14,6 +14,9 @@
 //===----------------------------------------------------------------------===//
 
 #include "AMDGPU.h"
+#include "AMDGPUTargetMachine.h"
+#include "Utils/AMDGPUBaseInfo.h"
+#include "llvm/ADT/SmallPtrSet.h"
 #include "llvm/IR/Module.h"
 #include "llvm/Transforms/Utils/Cloning.h"
 
@@ -30,13 +33,18 @@ static cl::opt<bool> StressCalls(
 class AMDGPUAlwaysInline : public ModulePass {
   bool GlobalOpt;
 
+  void recursivelyVisitUsers(GlobalValue &GV,
+                             SmallPtrSetImpl<Function *> &FuncsToAlwaysInline);
 public:
   static char ID;
 
   AMDGPUAlwaysInline(bool GlobalOpt = false) :
     ModulePass(ID), GlobalOpt(GlobalOpt) { }
   bool runOnModule(Module &M) override;
-  StringRef getPassName() const override { return "AMDGPU Always Inline Pass"; }
+
+  void getAnalysisUsage(AnalysisUsage &AU) const override {
+    AU.setPreservesAll();
+ }
 };
 
 } // End anonymous namespace
@@ -46,15 +54,53 @@ INITIALIZE_PASS(AMDGPUAlwaysInline, "amdgpu-always-inline",
 
 char AMDGPUAlwaysInline::ID = 0;
 
+void AMDGPUAlwaysInline::recursivelyVisitUsers(
+  GlobalValue &GV,
+  SmallPtrSetImpl<Function *> &FuncsToAlwaysInline) {
+  SmallVector<User *, 16> Stack;
+
+  SmallPtrSet<const Value *, 8> Visited;
+
+  for (User *U : GV.users())
+    Stack.push_back(U);
+
+  while (!Stack.empty()) {
+    User *U = Stack.pop_back_val();
+    if (!Visited.insert(U).second)
+      continue;
+
+    if (Instruction *I = dyn_cast<Instruction>(U)) {
+      Function *F = I->getParent()->getParent();
+      if (!AMDGPU::isEntryFunctionCC(F->getCallingConv())) {
+        FuncsToAlwaysInline.insert(F);
+        Stack.push_back(F);
+      }
+
+      // No need to look at further users, but we do need to inline any callers.
+      continue;
+    }
+
+    for (User *UU : U->users())
+      Stack.push_back(UU);
+  }
+}
+
 bool AMDGPUAlwaysInline::runOnModule(Module &M) {
+  AMDGPUAS AMDGPUAS = AMDGPU::getAMDGPUAS(M);
+
   std::vector<GlobalAlias*> AliasesToRemove;
-  std::vector<Function *> FuncsToClone;
+
+  SmallPtrSet<Function *, 8> FuncsToAlwaysInline;
+  SmallPtrSet<Function *, 8> FuncsToNoInline;
 
   for (GlobalAlias &A : M.aliases()) {
     if (Function* F = dyn_cast<Function>(A.getAliasee())) {
       A.replaceAllUsesWith(F);
       AliasesToRemove.push_back(&A);
     }
+
+    // FIXME: If the aliasee isn't a function, it's some kind of constant expr
+    // cast that won't be inlined through.
   }
 
   if (GlobalOpt) {
@@ -63,31 +109,51 @@ bool AMDGPUAlwaysInline::runOnModule(Module &M) {
     }
   }
 
-  auto NewAttr = StressCalls ? Attribute::NoInline : Attribute::AlwaysInline;
-  auto IncompatAttr
-    = StressCalls ? Attribute::AlwaysInline : Attribute::NoInline;
-
-  for (Function &F : M) {
-    if (!F.hasLocalLinkage() && !F.isDeclaration() && !F.use_empty() &&
-        !F.hasFnAttribute(IncompatAttr))
-      FuncsToClone.push_back(&F);
-  }
-
-  for (Function *F : FuncsToClone) {
-    ValueToValueMapTy VMap;
-    Function *NewFunc = CloneFunction(F, VMap);
-    NewFunc->setLinkage(GlobalValue::InternalLinkage);
-    F->replaceAllUsesWith(NewFunc);
+  // Always force inlining of any function that uses an LDS global address. This
+  // is something of a workaround because we don't have a way of supporting LDS
+  // objects defined in functions. LDS is always allocated by a kernel, and it
+  // is difficult to manage LDS usage if a function may be used by multiple
+  // kernels.
+  //
+  // OpenCL doesn't allow declaring LDS in non-kernels, so in practice this
+  // should only appear when IPO passes manages to move LDs defined in a kernel
+  // into a single user function.
+
+  for (GlobalVariable &GV : M.globals()) {
+    // TODO: Region address
+    unsigned AS = GV.getType()->getAddressSpace();
+    if (AS != AMDGPUAS::LOCAL_ADDRESS && AS != AMDGPUAS.REGION_ADDRESS)
+      continue;
+
+    recursivelyVisitUsers(GV, FuncsToAlwaysInline);
   }
 
-  for (Function &F : M) {
-    if (F.hasLocalLinkage() && !F.hasFnAttribute(IncompatAttr)) {
-      F.addFnAttr(NewAttr);
+  if (!AMDGPUTargetMachine::EnableFunctionCalls || StressCalls) {
+    auto IncompatAttr
+      = StressCalls ? Attribute::AlwaysInline : Attribute::NoInline;
+
+    for (Function &F : M) {
+      if (!F.isDeclaration() && !F.use_empty() &&
+          !F.hasFnAttribute(IncompatAttr)) {
+        if (StressCalls) {
+          if (!FuncsToAlwaysInline.count(&F))
+            FuncsToNoInline.insert(&F);
+        } else
+          FuncsToAlwaysInline.insert(&F);
+      }
     }
   }
-  return false;
+
+  for (Function *F : FuncsToAlwaysInline)
+    F->addFnAttr(Attribute::AlwaysInline);
+
+  for (Function *F : FuncsToNoInline)
+    F->addFnAttr(Attribute::NoInline);
+
+  return !FuncsToAlwaysInline.empty() || !FuncsToNoInline.empty();
 }
 
 ModulePass *llvm::createAMDGPUAlwaysInlinePass(bool GlobalOpt) {
   return new AMDGPUAlwaysInline(GlobalOpt);
 }
+
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp b/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp
index 22ea16049bf..b4066102553 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp
@@ -117,11 +117,12 @@ static cl::opt<bool, true> LateCFGStructurize(
   cl::location(AMDGPUTargetMachine::EnableLateStructurizeCFG),
   cl::Hidden);
 
-static cl::opt<bool> EnableAMDGPUFunctionCalls(
+static cl::opt<bool, true> EnableAMDGPUFunctionCalls(
   "amdgpu-function-calls",
-  cl::Hidden,
   cl::desc("Enable AMDGPU function call support"),
-  cl::init(false));
+  cl::location(AMDGPUTargetMachine::EnableFunctionCalls),
+  cl::init(false),
+  cl::Hidden);
 
 // Enable lib calls simplifications
 static cl::opt<bool> EnableLibCallSimplify(
@@ -311,9 +312,10 @@ AMDGPUTargetMachine::AMDGPUTargetMachine(const Target &T, const Triple &TT,
   initAsmInfo();
 }
 
-AMDGPUTargetMachine::~AMDGPUTargetMachine() = default;
-
 bool AMDGPUTargetMachine::EnableLateStructurizeCFG = false;
+bool AMDGPUTargetMachine::EnableFunctionCalls = false;
+
+AMDGPUTargetMachine::~AMDGPUTargetMachine() = default;
 
 StringRef AMDGPUTargetMachine::getGPUName(const Function &F) const {
   Attribute GPUAttr = F.getFnAttribute("target-cpu");
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.h b/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.h
index 4dcb1afd313..50b219d639c 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.h
+++ b/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.h
@@ -41,6 +41,7 @@ protected:
 
 public:
   static bool EnableLateStructurizeCFG;
+  static bool EnableFunctionCalls;
 
   AMDGPUTargetMachine(const Target &T, const Triple &TT, StringRef CPU,
                       StringRef FS, TargetOptions Options,